The Discovery Of Adverse Drug Reactions Based On Comment Mining

As the harm of adverse drug reactions (ADRs) grows, drug safety issues draws more people’s attention and is becoming the focus of the medical professionals and the public, so how to discover the ADRs is of great theoretical and practical value. With the development of Web2.0technologies, many healthcare social networking sites appear on the Internet and people share medication experiences and give comments on drugs there. As the comments on those websites become increasingly rich, researchers begin to pay attention to the ADR information within the user comments, and gradually develop a quick and effective ADR discovery mechanism through mining the comments.When mining adverse drug reactions (ADRs) from the comments, it is very important to recognize novel ADR expressions and normalize them, since people probably adopt different expressions to describe adverse reactions and new adverse reactions may emerge with the listing of new drugs as well as the diversity of drug users. For this case, the work in Chapter3utilized the conditional random field (CRF) model to recognize adverse reaction entities, and proposed a normalization method for them. Experimental results indicated that the CRF was able to identify both known and novel adverse reaction entities, and the normalization merged those entities, which benefitted the ADR discovery. The similarity between mined results of known ADRs and database records verified the effectiveness of this mining method, and a list of potential ADRs sorted by occurrence frequency in comments was obtained finally.The recognition of adverse reaction entities from user comments is a basic but crucial step in the discovery of ADRs. Because of the grammatical irregularity of the comments and the diversity of the adverse reaction expressions, it is rather challenging to recognize adverse reaction entities from the comments. To solve this problem, the work in Chapter4implemented an entity recognition framework which integrated the results of different recognition methods. The first method recognized entities by matching the bag of words within a sliding window with the word bag of lexical terms, where the edit distances of words were considered; the second one adopted the CRF model to recognize entities, which applied feature selection to find the best internal feature set, and identified the most effective feature combination through repeatedly trials. The recognition results of the two methods were integrated and the integrated performance is greatly improved compared with either result of the two recognition methods, indicating that the integration can compensate the deficiencies of a single recognition method. Compared with other adverse reaction entity recognition methods, the performance of this method is comparable or maybe even better than theirs, demonstrating the effectiveness of the proposed method.